ABSTRACT
Diffusion Models is a primary class of Generative AI methods that can be used to create synthetic images or videos. First, generative modeling is overviewed. Then competing Generative AI methods, such as Variational Auto-Encoder (VAE) or Generative Adversarial Networks (GAN) and their limitations are presented. The presentation of the primary Diffusion Models follow, namely, Denoising Diffusion Probabilistic Models (DDPMs), Score-based Generative Models (SGMs) and Score-based Generative Model with Stochastic Differential Equations (SDEs). Finally, Diffusion Model applications are overviewed, namely, on unconditional image generation, image super-resolution, inpainting, repainting, or restoration, semantic image segmentation, video generation and point cloud completion – generation.
Figure 1: Diffusion Model Process.
Figure 2: Image repainting.
Diffusion-Models-v3.0-summary